item_id stringlengths 14 17 | set stringclasses 2
values | session_tokens int64 38.9k 109k | budget_tokens int64 1.95k 3.07k | session stringlengths 145k 447k | haiku_summary stringlengths 355 25k | haiku_passed bool 2
classes | haiku_item_score float64 0 0.77 | haiku_nine_headings bool 2
classes | haiku_within_budget bool 2
classes | haiku_recall float64 0 1 | haiku_hallucination float64 0 0.17 | haiku_continuation float64 0 1 | haiku_cost_usd float64 0.01 0.15 | haiku_time_s float64 21.9 90.6 | haiku_output_tokens int64 5.14k 20k | sonnet_summary stringlengths 9.62k 27k | sonnet_passed bool 2
classes | sonnet_item_score float64 0 0.8 | sonnet_nine_headings bool 1
class | sonnet_within_budget bool 2
classes | sonnet_recall float64 0.46 1 | sonnet_hallucination float64 0 0.21 | sonnet_continuation float64 0.25 0.75 | sonnet_cost_usd float64 0.19 0.46 | sonnet_time_s float64 23.5 77.6 | sonnet_output_tokens int64 3.95k 10.9k | microtensor_summary stringlengths 699 7.45k | microtensor_passed bool 2
classes | microtensor_item_score float64 0.13 0.93 | microtensor_nine_headings bool 1
class | microtensor_within_budget bool 1
class | microtensor_recall float64 0 1 | microtensor_hallucination float64 0 0.1 | microtensor_continuation float64 0.25 1 | microtensor_cost_usd float64 0.01 0.01 | microtensor_time_s float64 14.1 51.4 | microtensor_output_tokens int64 166 1.9k |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
exp1-a-s1-s000 | A | 54,317 | 2,715 | "<session>\n[user]\nPicking up FIN-2319: the nightly ledger export has failed every night since Mond(...TRUNCATED) | "<analysis>\nThis summary covers FIN-2319 (nightly export failures), FIN-2871 (statement export Temp(...TRUNCATED) | false | 0 | false | true | 0.666667 | 0 | 0.5 | 0.019276 | 84.0283 | 20,000 | "<analysis>\nChronological review of the conversation, with attention to accuracy. Several of my ear(...TRUNCATED) | false | 0.535714 | true | true | 0.571429 | 0 | 0.5 | 0.269452 | 53.3284 | 8,393 | "<analysis>\nSummary of the conversation so far, in the nine requested sections.\n</analysis>\n\n<su(...TRUNCATED) | false | 0.556818 | true | true | 0.666667 | 0.045455 | 0.5 | 0.005 | 37.6995 | 483 |
exp1-a-s1-s001 | A | 63,265 | 3,072 | "<session>\n[user]\nPicking up FIN-2817: the nightly ledger export has failed every night since Mond(...TRUNCATED) | "<analysis>\nThis summary covers the FIN-2817 nightly export failure work and the many follow-on req(...TRUNCATED) | false | 0 | true | false | 0.789474 | 0.055556 | 0.5 | 0.091436 | 61.5723 | 15,288 | "<analysis>\nChronological review of the session, including places where my own earlier statements w(...TRUNCATED) | false | 0.454076 | true | true | 0.631579 | 0.058824 | 0.333333 | 0.271386 | 40.6446 | 5,852 | "<analysis>\nSummary of the conversation so far, in the nine requested sections.\n</analysis>\n\n<su(...TRUNCATED) | false | 0.432133 | true | true | 0.578947 | 0.052632 | 0.333333 | 0.005 | 39.4809 | 381 |
exp1-a-s1-s002 | A | 63,862 | 3,072 | "<session>\n[user]\nCI on perf/lockfile-warmup-4 (PR #14629) keeps going red. Start with the e2e fla(...TRUNCATED) | "<analysis>\nThis is a long, multi-thread CI/Kubernetes/TypeScript session. The user's most recent r(...TRUNCATED) | false | 0 | true | false | 0.923077 | 0.04386 | 0.5 | 0.097693 | 70.7822 | 17,440 | "<analysis>\nChronological review of the conversation, noting where my earlier turns were inaccurate(...TRUNCATED) | true | 0.674089 | true | true | 0.923077 | 0.052632 | 0.5 | 0.319142 | 69.7845 | 10,277 | "<analysis>\nSummary of the conversation so far, in the nine requested sections.\n</analysis>\n\n<su(...TRUNCATED) | true | 0.798077 | true | true | 0.846154 | 0 | 0.75 | 0.005 | 39.7684 | 585 |
exp1-a-s1-s003 | A | 69,140 | 3,072 | "<session>\n[user]\nCI on perf/shard-warmup-9 (PR #14586) keeps going red. Start with the e2e flake,(...TRUNCATED) | "<analysis>\nThis is a long CI-debugging session on the atlas repo (PR #14586 on perf/shard-warmup-9(...TRUNCATED) | false | 0 | false | true | 0 | 0 | 0.25 | 0.108576 | 85.9528 | 20,000 | "<analysis>\nChronological review of the conversation, with attention to where my own earlier statem(...TRUNCATED) | true | 0.672009 | true | true | 0.923077 | 0.055556 | 0.5 | 0.335152 | 77.5509 | 10,085 | "<analysis>\nSummary of the conversation so far, in the nine requested sections.\n</analysis>\n\n<su(...TRUNCATED) | false | 0.125 | true | true | 0 | 0 | 0.25 | 0.005 | 40.7228 | 222 |
exp1-a-s1-s004 | A | 62,236 | 3,072 | "<session>\n[user]\nThe orders migration keeps failing on staging. Can you find out why? Attempt log(...TRUNCATED) | "<analysis>\nThis conversation covers a staging migration investigation plus a long series of smalle(...TRUNCATED) | false | 0 | false | true | 0.857143 | 0 | 0.75 | 0.101736 | 82.977 | 20,000 | "<analysis>\nChronological review of the conversation:\n\n1. The user asked why the orders migration(...TRUNCATED) | true | 0.803571 | true | true | 0.857143 | 0 | 0.75 | 0.290712 | 49.5064 | 8,377 | "<analysis>\nSummary of the conversation so far, in the nine requested sections.\n</analysis>\n\n<su(...TRUNCATED) | true | 0.928571 | true | true | 0.857143 | 0 | 1 | 0.005 | 20.3921 | 1,191 |
exp1-a-s1-s005 | A | 58,016 | 2,900 | "<session>\n[user]\nThe orders migration keeps failing on staging. Can you find out why? Attempt log(...TRUNCATED) | "<analysis>\nThis conversation covers a long run of work on the orders migration on staging, followe(...TRUNCATED) | false | 0 | false | true | 0.263158 | 0 | 0.4 | 0.019733 | 85.0779 | 20,000 | "<analysis>\nChronological review of the conversation, noting where my own earlier reports were inac(...TRUNCATED) | false | 0.596546 | true | true | 0.631579 | 0.03125 | 0.6 | 0.270846 | 47.7015 | 7,619 | "<analysis>\nSummary of the conversation so far, in the nine requested sections.\n</analysis>\n\n<su(...TRUNCATED) | true | 0.689474 | true | true | 0.578947 | 0 | 0.8 | 0.005 | 20.6933 | 1,268 |
exp1-a-s1-s006 | A | 71,346 | 3,072 | "<session>\n[user]\nTwo things on the relay gateway. Security flagged CVE-2026-46096 in ws, so we ha(...TRUNCATED) | "<analysis>\nThe conversation covers two threads. The main one is the relay gateway work: a ws CVE u(...TRUNCATED) | false | 0 | true | false | 0.923077 | 0.039216 | 0.25 | 0.109466 | 85.2953 | 20,000 | "<analysis>\nChronological review of the session, including where my earlier statements were wrong o(...TRUNCATED) | true | 0.679196 | true | true | 0.923077 | 0.045455 | 0.5 | 0.346594 | 75.4237 | 10,873 | "<analysis>\nSummary of the conversation so far, in the nine requested sections.\n</analysis>\n\n<su(...TRUNCATED) | true | 0.8125 | true | true | 1 | 0.071429 | 0.75 | 0.005 | 40.6114 | 524 |
exp1-a-s1-s007 | A | 61,549 | 3,072 | "<session>\n[user]\nTwo things on the relay gateway. Security flagged CVE-2026-37093 in ws, so we ha(...TRUNCATED) | "<analysis>\nThis is a summary request, so no tools. Reviewing the session, several of my earlier re(...TRUNCATED) | false | 0 | false | false | 0.923077 | 0.075 | 0.5 | 0.101666 | 83.7976 | 20,000 | "<analysis>\nChronological review of the conversation, including places where my earlier turns were (...TRUNCATED) | true | 0.646853 | true | true | 0.923077 | 0.090909 | 0.5 | 0.295602 | 65.0385 | 8,894 | "<analysis>\nSummary of the conversation so far, in the nine requested sections.\n</analysis>\n\n<su(...TRUNCATED) | false | 0.125 | true | true | 0 | 0 | 0.25 | 0.005 | 39.1953 | 466 |
exp1-a-s1-s008 | A | 59,800 | 2,990 | "<session>\n[user]\nHanding this over to you: checkout error rates jumped after the v3.51.0 deploy a(...TRUNCATED) | "<analysis>\nThis conversation covers an incident rollback for checkout v3.51.0 plus a long run of s(...TRUNCATED) | false | 0 | false | true | 0.705882 | 0 | 0.2 | 0.10095 | 80.7009 | 20,000 | "<analysis>\nChronological review of the conversation, with attention to what was actually done vers(...TRUNCATED) | false | 0.552941 | true | true | 0.705882 | 0 | 0.4 | 0.30929 | 64.6769 | 10,549 | "<analysis>\nSummary of the conversation so far, in the nine requested sections.\n</analysis>\n\n<su(...TRUNCATED) | false | 0.258824 | true | true | 0.117647 | 0 | 0.4 | 0.005 | 38.8584 | 407 |
exp1-a-s1-s009 | A | 61,233 | 3,061 | "<session>\n[user]\nHanding this over to you: checkout error rates jumped after the v3.52.0 deploy a(...TRUNCATED) | "<analysis>\nThis is a long, multi-request session. The main goal is a clean rollback of checkout fr(...TRUNCATED) | false | 0 | false | true | 0 | 0 | 0.2 | 0.102328 | 85.5005 | 20,000 | "<analysis>\nChronological review of the conversation, with attention to where my earlier turns were(...TRUNCATED) | false | 0.514943 | true | true | 0.666667 | 0.034483 | 0.4 | 0.277982 | 45.3759 | 6,867 | "<analysis>\nSummary of the conversation so far, in the nine requested sections.\n</analysis>\n\n<su(...TRUNCATED) | false | 0.485333 | true | true | 0.611111 | 0.04 | 0.4 | 0.005 | 39.0429 | 754 |
End of preview. Expand in Data Studio
Ditto pilot results
Replay bench of the Ditto router pilot. Each of the 36 sessions from Woolford's compaction study replication package (MIT) was turned into the compaction request Claude Code sends and passed through a fork of Ditto's router three ways: to Claude Haiku 5.5 and Claude Sonnet 5.5 as the default compaction model, and with the archetype route table on, which sends compaction to Microtensor's certified mt/ctx_compress.
compaction has one row per session with the session, the three summaries side by side, the arena grader's marks, the cost and the time. titles has the title request for each session; the Microtensor title route is a clearly labelled stand in (Qwen2.5 1.5B Instruct), not a certified model.
Code, relay logs and the method: https://github.com/microtensor-io/ditto-pilot
Charts
Numbers: cost_per_compaction.csv
Numbers: quality_vs_cost.csv
Numbers: time_per_compaction.csv
Numbers: requests_by_type_and_route.csv
Numbers: quality_by_measure.csv
Numbers: session_scores.csv
- Downloads last month
- 2